利用调查和人口普查数据评估小面积贫困估计的替代方法

Evaluating Alternative Approaches to SmallArea Estimation of Poverty with Surveyand Census Data Hai-Anh DangMinh DoPartha LahiriMelany GualavisiDavid NewhouseTalip KilicPeter LanjouwRoy Van der Weide Development EconomicsDevelopment Data GroupMay 2026Public Disclosure Authorized A verified reproducibility package for this paper isavailable athttp://reproducibility.worldbank.org,clickherefor direct access. Policy Research Working Paper11396 Abstract This paper uses five rounds of Mexican and Brazilian censusextracts to evaluate the accuracy of different model specifi-cations and estimation methods that use survey and censusdata to generate small area estimates of poverty. Models thatutilize more granular data for prediction (household- and/or village-level predictors) tend to pro-duce more accurateestimates of poverty than models estimated only using area-level predictors. Differences in accuracy across models andmethods that utilize household or village level predictorsare minor. Models that omit household-level predictorstend to be more robust than unit-level models to the use of old census data and classical measurement error in surveypredictors. The performance of the Fay-Herriot area-levelmodel falls in the presence of sample selection bias andsmall sample sizes. Rescaling sample weights is importantin Mexico, where the sample is informative within areas.Applying raw sample weights without rescaling in this casegreatly reduces the accuracy of estimates from linear modelsand distorts methodological comparisons. Overall, no oneapproach dominates across all contexts, but when sampleweights are rescaled there is no downside to using moregranular data for prediction. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about developmentissues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry thenames of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely thoseof the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank andits affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Evaluating alternative approaches to small areaestimation of poverty with survey and census data Hai-Anh Dangr○∗Minh Dor○∗Partha Lahirir○†Melany Gualavisir○‡David Newhouser○∗Talip Kilicr○∗Peter Lanjouwr○§Roy Van der Weider○¶ 1Introduction Small area estimation (SAE) enables survey data to "borrow strength" from more geo-graphically comprehensive auxiliary data such as census data. This enables the estimationof survey-based indicators such as poverty for highly disaggregated areas or subgroups forwhich there are no or insufficient household survey data to obtain reliable direct estimates.SAE can therefore be a critical input into the geographic targeting and evaluation of policiesand programs. The first known application of the SAE of poverty, combining survey and census data,was for small places in the US (Fay and Herriot, 1979) and employed Empirical Bayesestimation (Carter and Rolph (1974) and Efron and Morris (1977)). The US Census Bureausubsequently established the Small Area Income and Poverty Estimates program, whichregularly produces small area estimates of income and poverty for school districts andcounties.Later, an alternative method that combines survey and census data to obtainsmall area poverty estimates was developed by Elbers et al. (2003), henceforward referred toas the ELL approach.1This method relaxes distributional assumptions regarding the errorterm and employs Monte-Carlo simulation techniques advocated by Berry et al. (1995). Ithas been used in over 60 countries worldwide, often with the support of the World Bank,as well as in several notable research applications (Demombynes and Özler (2005), Elberset al. (2005), Elbers et al. (2007), Araujo et al. (2008), Andam et al. (2010), Fujii (2010),Crost et al. (2014), Enamorado et al. (2016), Bazzi (2017)). SAE models can broadly be divided into three groups depending on the level at whichthey are specified:Area-level modelsspecified at the target area level,unit-level modelsspecified at the household level (which represents the most disaggregated level), andsub-area level modelsspecified at an aggregate level that is below the target area level, such asthe village level. Any of these models could in principle be used to obtain purely synthetic predictions, as in Elbers et al. (2003), or alternatively to obtain Empirical Best Predictors(EBP) that incorporate a random effect that is conditioned on the available survey sampledata (Laird and Ware, 1982; Battese et al., 1988; Jiang and Lahiri, 2006; Molina and Rao,2010; Elbers and Van der Weide, 2026). When working with unit or sub-area-level models,